| Message ID | 1476147473-30970-8-git-send-email-feng.wu@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
> From: Wu, Feng > Sent: Tuesday, October 11, 2016 8:58 AM > > When cpu is offline, we need to move all the vcpus in its blocking > list to another online cpu, this patch handles it. > > Signed-off-by: Feng Wu <feng.wu@intel.com> > --- > v5: > - Add some comments to explain why it doesn't cause deadlock > for the ABBA deadlock scenario. > > xen/arch/x86/hvm/vmx/vmcs.c | 1 + > xen/arch/x86/hvm/vmx/vmx.c | 48 > +++++++++++++++++++++++++++++++++++++++ > xen/include/asm-x86/hvm/vmx/vmx.h | 1 + > 3 files changed, 50 insertions(+) > > diff --git a/xen/arch/x86/hvm/vmx/vmcs.c b/xen/arch/x86/hvm/vmx/vmcs.c > index 10976bd..5dd68ca 100644 > --- a/xen/arch/x86/hvm/vmx/vmcs.c > +++ b/xen/arch/x86/hvm/vmx/vmcs.c > @@ -578,6 +578,7 @@ void vmx_cpu_dead(unsigned int cpu) > vmx_free_vmcs(per_cpu(vmxon_region, cpu)); > per_cpu(vmxon_region, cpu) = 0; > nvmx_cpu_dead(cpu); > + vmx_pi_desc_fixup(cpu); > } > > int vmx_cpu_up(void) > diff --git a/xen/arch/x86/hvm/vmx/vmx.c b/xen/arch/x86/hvm/vmx/vmx.c > index b14c84e..c71d496 100644 > --- a/xen/arch/x86/hvm/vmx/vmx.c > +++ b/xen/arch/x86/hvm/vmx/vmx.c > @@ -208,6 +208,54 @@ static void vmx_pi_do_resume(struct vcpu *v) > vmx_pi_list_remove(v); > } > > +void vmx_pi_desc_fixup(int cpu) > +{ > + unsigned int new_cpu, dest; > + unsigned long flags; > + struct arch_vmx_struct *vmx, *tmp; > + spinlock_t *new_lock, *old_lock = &per_cpu(vmx_pi_blocking, cpu).lock; > + struct list_head *blocked_vcpus = &per_cpu(vmx_pi_blocking, cpu).list; > + > + if ( !iommu_intpost ) > + return; > + > + /* > + * We are in the context of CPU_DEAD or CPU_UP_CANCELED notification, > + * and it is impossible for a second CPU go down in parallel. So we > + * can safely acquire the old cpu's lock and then acquire the new_cpu's > + * lock after that. > + */ > + spin_lock_irqsave(old_lock, flags); > + > + list_for_each_entry_safe(vmx, tmp, blocked_vcpus, pi_blocking.list) > + { > + /* > + * We need to find an online cpu as the NDST of the PI descriptor, it > + * doesn't matter whether it is within the cpupool of the domain or > + * not. As long as it is online, the vCPU will be woken up once the > + * notification event arrives. > + */ > + new_cpu = cpumask_any(&cpu_online_map); > + new_lock = &per_cpu(vmx_pi_blocking, new_cpu).lock; > + > + spin_lock(new_lock); > + > + ASSERT(vmx->pi_blocking.lock == old_lock); > + > + dest = cpu_physical_id(new_cpu); > + write_atomic(&vmx->pi_desc.ndst, > + x2apic_enabled ? dest : MASK_INSR(dest, PI_xAPIC_NDST_MASK)); > + > + list_move(&vmx->pi_blocking.list, > + &per_cpu(vmx_pi_blocking, new_cpu).list); > + vmx->pi_blocking.lock = new_lock; > + > + spin_unlock(new_lock); I didn't check the whole flow... but did you suppress notification somewhere earlier before above list movement happens? Otherwise you may miss an interrupt when the target cpu is dying... > + } > + > + spin_unlock_irqrestore(old_lock, flags); > +} > + > /* This function is called when pcidevs_lock is held */ > void vmx_pi_hooks_assign(struct domain *d) > { > diff --git a/xen/include/asm-x86/hvm/vmx/vmx.h > b/xen/include/asm-x86/hvm/vmx/vmx.h > index 4cdd9b1..9783c70 100644 > --- a/xen/include/asm-x86/hvm/vmx/vmx.h > +++ b/xen/include/asm-x86/hvm/vmx/vmx.h > @@ -569,6 +569,7 @@ void free_p2m_hap_data(struct p2m_domain *p2m); > void p2m_init_hap_data(struct p2m_domain *p2m); > > void vmx_pi_per_cpu_init(unsigned int cpu); > +void vmx_pi_desc_fixup(int cpu); > > void vmx_pi_hooks_assign(struct domain *d); > void vmx_pi_hooks_deassign(struct domain *d); > -- > 2.1.0
> -----Original Message----- > From: Tian, Kevin > Sent: Tuesday, October 11, 2016 4:38 PM > To: Wu, Feng <feng.wu@intel.com>; xen-devel@lists.xen.org > Cc: jbeulich@suse.com; andrew.cooper3@citrix.com; > george.dunlap@eu.citrix.com; dario.faggioli@citrix.com > Subject: RE: [PATCH v5 7/7] VMX: Fixup PI descriptor when cpu is offline > > > From: Wu, Feng > > Sent: Tuesday, October 11, 2016 8:58 AM > > > > When cpu is offline, we need to move all the vcpus in its blocking > > list to another online cpu, this patch handles it. > > > > Signed-off-by: Feng Wu <feng.wu@intel.com> > > --- > > v5: > > - Add some comments to explain why it doesn't cause deadlock > > for the ABBA deadlock scenario. > > > > xen/arch/x86/hvm/vmx/vmcs.c | 1 + > > xen/arch/x86/hvm/vmx/vmx.c | 48 > > +++++++++++++++++++++++++++++++++++++++ > > xen/include/asm-x86/hvm/vmx/vmx.h | 1 + > > 3 files changed, 50 insertions(+) > > > > diff --git a/xen/arch/x86/hvm/vmx/vmcs.c b/xen/arch/x86/hvm/vmx/vmcs.c > > index 10976bd..5dd68ca 100644 > > --- a/xen/arch/x86/hvm/vmx/vmcs.c > > +++ b/xen/arch/x86/hvm/vmx/vmcs.c > > @@ -578,6 +578,7 @@ void vmx_cpu_dead(unsigned int cpu) > > vmx_free_vmcs(per_cpu(vmxon_region, cpu)); > > per_cpu(vmxon_region, cpu) = 0; > > nvmx_cpu_dead(cpu); > > + vmx_pi_desc_fixup(cpu); > > } > > > > int vmx_cpu_up(void) > > diff --git a/xen/arch/x86/hvm/vmx/vmx.c b/xen/arch/x86/hvm/vmx/vmx.c > > index b14c84e..c71d496 100644 > > --- a/xen/arch/x86/hvm/vmx/vmx.c > > +++ b/xen/arch/x86/hvm/vmx/vmx.c > > @@ -208,6 +208,54 @@ static void vmx_pi_do_resume(struct vcpu *v) > > vmx_pi_list_remove(v); > > } > > > > +void vmx_pi_desc_fixup(int cpu) > > +{ > > + unsigned int new_cpu, dest; > > + unsigned long flags; > > + struct arch_vmx_struct *vmx, *tmp; > > + spinlock_t *new_lock, *old_lock = &per_cpu(vmx_pi_blocking, cpu).lock; > > + struct list_head *blocked_vcpus = &per_cpu(vmx_pi_blocking, cpu).list; > > + > > + if ( !iommu_intpost ) > > + return; > > + > > + /* > > + * We are in the context of CPU_DEAD or CPU_UP_CANCELED notification, > > + * and it is impossible for a second CPU go down in parallel. So we > > + * can safely acquire the old cpu's lock and then acquire the new_cpu's > > + * lock after that. > > + */ > > + spin_lock_irqsave(old_lock, flags); > > + > > + list_for_each_entry_safe(vmx, tmp, blocked_vcpus, pi_blocking.list) > > + { > > + /* > > + * We need to find an online cpu as the NDST of the PI descriptor, it > > + * doesn't matter whether it is within the cpupool of the domain or > > + * not. As long as it is online, the vCPU will be woken up once the > > + * notification event arrives. > > + */ > > + new_cpu = cpumask_any(&cpu_online_map); > > + new_lock = &per_cpu(vmx_pi_blocking, new_cpu).lock; > > + > > + spin_lock(new_lock); > > + > > + ASSERT(vmx->pi_blocking.lock == old_lock); > > + > > + dest = cpu_physical_id(new_cpu); > > + write_atomic(&vmx->pi_desc.ndst, > > + x2apic_enabled ? dest : MASK_INSR(dest, PI_xAPIC_NDST_MASK)); > > + > > + list_move(&vmx->pi_blocking.list, > > + &per_cpu(vmx_pi_blocking, new_cpu).list); > > + vmx->pi_blocking.lock = new_lock; > > + > > + spin_unlock(new_lock); > > I didn't check the whole flow... but did you suppress notification somewhere > earlier before above list movement happens? Otherwise you may miss an > interrupt when the target cpu is dying... > Yes, this is really a good point. the wakeup notification event will miss if it comes in before " write_atomic(&vmx->pi_desc.ndst ...." above. So we need to suppress it before "spin_lock_irqsave(old_lock, flags);". Thanks, Feng
diff --git a/xen/arch/x86/hvm/vmx/vmcs.c b/xen/arch/x86/hvm/vmx/vmcs.c index 10976bd..5dd68ca 100644 --- a/xen/arch/x86/hvm/vmx/vmcs.c +++ b/xen/arch/x86/hvm/vmx/vmcs.c @@ -578,6 +578,7 @@ void vmx_cpu_dead(unsigned int cpu) vmx_free_vmcs(per_cpu(vmxon_region, cpu)); per_cpu(vmxon_region, cpu) = 0; nvmx_cpu_dead(cpu); + vmx_pi_desc_fixup(cpu); } int vmx_cpu_up(void) diff --git a/xen/arch/x86/hvm/vmx/vmx.c b/xen/arch/x86/hvm/vmx/vmx.c index b14c84e..c71d496 100644 --- a/xen/arch/x86/hvm/vmx/vmx.c +++ b/xen/arch/x86/hvm/vmx/vmx.c @@ -208,6 +208,54 @@ static void vmx_pi_do_resume(struct vcpu *v) vmx_pi_list_remove(v); } +void vmx_pi_desc_fixup(int cpu) +{ + unsigned int new_cpu, dest; + unsigned long flags; + struct arch_vmx_struct *vmx, *tmp; + spinlock_t *new_lock, *old_lock = &per_cpu(vmx_pi_blocking, cpu).lock; + struct list_head *blocked_vcpus = &per_cpu(vmx_pi_blocking, cpu).list; + + if ( !iommu_intpost ) + return; + + /* + * We are in the context of CPU_DEAD or CPU_UP_CANCELED notification, + * and it is impossible for a second CPU go down in parallel. So we + * can safely acquire the old cpu's lock and then acquire the new_cpu's + * lock after that. + */ + spin_lock_irqsave(old_lock, flags); + + list_for_each_entry_safe(vmx, tmp, blocked_vcpus, pi_blocking.list) + { + /* + * We need to find an online cpu as the NDST of the PI descriptor, it + * doesn't matter whether it is within the cpupool of the domain or + * not. As long as it is online, the vCPU will be woken up once the + * notification event arrives. + */ + new_cpu = cpumask_any(&cpu_online_map); + new_lock = &per_cpu(vmx_pi_blocking, new_cpu).lock; + + spin_lock(new_lock); + + ASSERT(vmx->pi_blocking.lock == old_lock); + + dest = cpu_physical_id(new_cpu); + write_atomic(&vmx->pi_desc.ndst, + x2apic_enabled ? dest : MASK_INSR(dest, PI_xAPIC_NDST_MASK)); + + list_move(&vmx->pi_blocking.list, + &per_cpu(vmx_pi_blocking, new_cpu).list); + vmx->pi_blocking.lock = new_lock; + + spin_unlock(new_lock); + } + + spin_unlock_irqrestore(old_lock, flags); +} + /* This function is called when pcidevs_lock is held */ void vmx_pi_hooks_assign(struct domain *d) { diff --git a/xen/include/asm-x86/hvm/vmx/vmx.h b/xen/include/asm-x86/hvm/vmx/vmx.h index 4cdd9b1..9783c70 100644 --- a/xen/include/asm-x86/hvm/vmx/vmx.h +++ b/xen/include/asm-x86/hvm/vmx/vmx.h @@ -569,6 +569,7 @@ void free_p2m_hap_data(struct p2m_domain *p2m); void p2m_init_hap_data(struct p2m_domain *p2m); void vmx_pi_per_cpu_init(unsigned int cpu); +void vmx_pi_desc_fixup(int cpu); void vmx_pi_hooks_assign(struct domain *d); void vmx_pi_hooks_deassign(struct domain *d);
When cpu is offline, we need to move all the vcpus in its blocking list to another online cpu, this patch handles it. Signed-off-by: Feng Wu <feng.wu@intel.com> --- v5: - Add some comments to explain why it doesn't cause deadlock for the ABBA deadlock scenario. xen/arch/x86/hvm/vmx/vmcs.c | 1 + xen/arch/x86/hvm/vmx/vmx.c | 48 +++++++++++++++++++++++++++++++++++++++ xen/include/asm-x86/hvm/vmx/vmx.h | 1 + 3 files changed, 50 insertions(+)